Partial-client-participation analysis for personalized federated minimax optimization

Establish convergence guarantees and suitable techniques for extending Personalized Projected Ascent and Personalized VR-SGDA to partial client participation, where inactive clients may resume with stale personalized ascent variables while the shared model has evolved across communication rounds.

Background

The paper analyzes personalized federated minimax optimization under full client participation: each client maintains a private maximization variable while the shared minimization model is periodically aggregated. The convergence analysis relies on the alignment between each client’s local personalized variable and the evolving shared model.

Under partial participation, a dormant client retains an outdated inner variable and, when sampled again, must update it using a shared model that may have substantially drifted during the client’s absence. The authors identify this stale-inner-variable/global-model mismatch as a potential obstacle to the tracking bounds established for full participation and leave the extension unresolved.

References

Finally, this solution works perfectly when all clients participate. In the partial participation setting when inactive clients are sampled, their outdated local variables severely clash with the newly updated global model, potentially breaking the current techniques and the extension remains open.

— Personalized Projected Ascent for Federated Stackelberg Optimization  (2610.21295 - Yashwanth et al., 15 Oct 2026) in Conclusion (also discussed in the paragraph beginning “Scope” before Section 4)